StageNet/model.py at master · bruicecode/StageNet · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
import torch
from torch import nn
import torch.nn.utils.rnn as rnn_utils
from torch.utils import data
from torch.autograd import Variable
import torch.nn.functional as F

RANDOM_SEED = 12345
torch.manual_seed(RANDOM_SEED)
torch.cuda.manual_seed(RANDOM_SEED)
torch.backends.cudnn.deterministic=True

class StageNet(nn.Module):
    def __init__(self, input_dim, hidden_dim, conv_size, output_dim, levels, dropconnect=0., dropout=0., dropres=0.3):
        super(StageNet, self).__init__()

        assert hidden_dim % levels == 0
        self.dropout = dropout
        self.dropconnect = dropconnect
        self.dropres = dropres
        self.input_dim = input_dim
        self.hidden_dim = hidden_dim
        self.conv_dim = hidden_dim
        self.conv_size = conv_size
        self.output_dim = output_dim
        self.levels = levels
        self.chunk_size = hidden_dim // levels

        self.kernel = nn.Linear(int(input_dim+1), int(hidden_dim*4+levels*2))
        nn.init.xavier_uniform_(self.kernel.weight)
        nn.init.zeros_(self.kernel.bias)
        self.recurrent_kernel = nn.Linear(int(hidden_dim+1), int(hidden_dim*4+levels*2))
        nn.init.orthogonal_(self.recurrent_kernel.weight)
        nn.init.zeros_(self.recurrent_kernel.bias)

        self.nn_scale = nn.Linear(int(hidden_dim), int(hidden_dim // 6))
        self.nn_rescale = nn.Linear(int(hidden_dim // 6), int(hidden_dim))
        self.nn_conv = nn.Conv1d(int(hidden_dim), int(self.conv_dim), int(conv_size), 1)
        self.nn_output = nn.Linear(int(self.conv_dim), int(output_dim))

        if self.dropconnect:
            self.nn_dropconnect = nn.Dropout(p=dropconnect)
            self.nn_dropconnect_r = nn.Dropout(p=dropconnect)
        if self.dropout:
            self.nn_dropout = nn.Dropout(p=dropout)
            self.nn_dropres = nn.Dropout(p=dropres)

    def cumax(self, x, mode='l2r'):
        if mode == 'l2r':
            x = torch.softmax(x, dim=-1)
            x = torch.cumsum(x, dim=-1)
            return x
        elif mode == 'r2l':
            x = torch.flip(x, [-1])
            x = torch.softmax(x, dim=-1)
            x = torch.cumsum(x, dim=-1)
            return torch.flip(x, [-1])
        else:
            return x

    def step(self, inputs, c_last, h_last, interval):
        x_in = inputs

        # Integrate inter-visit time intervals
        interval = interval.unsqueeze(-1)
        x_out1 = self.kernel(torch.cat((x_in, interval), dim=-1))
        x_out2 = self.recurrent_kernel(torch.cat((h_last, interval), dim=-1))

        if self.dropconnect:
            x_out1 = self.nn_dropconnect(x_out1)
            x_out2 = self.nn_dropconnect_r(x_out2)
        x_out = x_out1 + x_out2
        f_master_gate = self.cumax(x_out[:, :self.levels], 'l2r')
        f_master_gate = f_master_gate.unsqueeze(2)
        i_master_gate = self.cumax(x_out[:, self.levels:self.levels*2], 'r2l')
        i_master_gate = i_master_gate.unsqueeze(2)
        x_out = x_out[:, self.levels*2:]
        x_out = x_out.reshape(-1, self.levels*4, self.chunk_size)
        f_gate = torch.sigmoid(x_out[:, :self.levels])
        i_gate = torch.sigmoid(x_out[:, self.levels:self.levels*2])
        o_gate = torch.sigmoid(x_out[:, self.levels*2:self.levels*3])
        c_in = torch.tanh(x_out[:, self.levels*3:])
        c_last = c_last.reshape(-1, self.levels, self.chunk_size)
        overlap = f_master_gate * i_master_gate
        c_out = overlap * (f_gate * c_last + i_gate * c_in) + (f_master_gate - overlap) * c_last + (i_master_gate - overlap) * c_in
        h_out = o_gate * torch.tanh(c_out)
        c_out = c_out.reshape(-1, self.hidden_dim)
        h_out = h_out.reshape(-1, self.hidden_dim)
        out = torch.cat([h_out, f_master_gate[..., 0], i_master_gate[..., 0]], 1)
        return out, c_out, h_out


    def forward(self, input, time, device):
        batch_size, time_step, feature_dim = input.size()
        c_out = torch.zeros(batch_size, self.hidden_dim).to(device)
        h_out = torch.zeros(batch_size, self.hidden_dim).to(device)

        tmp_h = torch.zeros_like(h_out, dtype=torch.float32).view(-1).repeat(self.conv_size).view(self.conv_size, batch_size, self.hidden_dim).to(device)
        tmp_dis = torch.zeros((self.conv_size, batch_size)).to(device)
        h = []
        origin_h = []
        distance = []
        for t in range(time_step):
            out, c_out, h_out = self.step(input[:, t, :], c_out, h_out, time[:, t])
            cur_distance = 1 - torch.mean(out[..., self.hidden_dim:self.hidden_dim+self.levels], -1)
            cur_distance_in = torch.mean(out[..., self.hidden_dim+self.levels:], -1)
            origin_h.append(out[..., :self.hidden_dim])
            tmp_h = torch.cat((tmp_h[1:], out[..., :self.hidden_dim].unsqueeze(0)), 0)
            tmp_dis = torch.cat((tmp_dis[1:], cur_distance.unsqueeze(0)), 0)
            distance.append(cur_distance)

            #Re-weighted convolution operation
            local_dis = tmp_dis.permute(1, 0)
            local_dis = torch.cumsum(local_dis, dim=1)
            local_dis = torch.softmax(local_dis, dim=1)
            local_h = tmp_h.permute(1, 2, 0)
            local_h = local_h * local_dis.unsqueeze(1)

            #Re-calibrate Progression patterns
            local_theme = torch.mean(local_h, dim=-1)
            local_theme = self.nn_scale(local_theme)
            local_theme = torch.relu(local_theme)
            local_theme = self.nn_rescale(local_theme)
            local_theme = torch.sigmoid(local_theme)

            local_h = self.nn_conv(local_h).squeeze(-1)
            local_h = local_theme * local_h
            h.append(local_h)

        origin_h = torch.stack(origin_h).permute(1, 0, 2)
        rnn_outputs = torch.stack(h).permute(1, 0, 2)
        if self.dropres > 0.0:
            origin_h = self.nn_dropres(origin_h)
        rnn_outputs = rnn_outputs + origin_h
        rnn_outputs = rnn_outputs.contiguous().view(-1, rnn_outputs.size(-1))
        if self.dropout > 0.0:
            rnn_outputs = self.nn_dropout(rnn_outputs)
        output = self.nn_output(rnn_outputs)
        output = output.contiguous().view(batch_size, time_step, self.output_dim)
        output = torch.sigmoid(output)

        return output, torch.stack(distance)